Today there are AI programs that can drive a car, beat the world go champion and write summaries of medical research papers. There are also industrial robots that can work very efficiently in precisely designed environments. These successes are largely due to the relatively new technology of deep learning, which is based on artificial nervous systems.

Despite many impressive advances, deep learning and current AI techniques have major limitations. Problems that are simple for humans are often impossible for today's AI programs. For example, there are no household robots today that can help with dishes, laundry and cleaning in our homes. Building such robots requires a radical increase in the adaptability of AI systems. One difficulty in this context is that deep learning generally uses fixed architectures that must be tailored by programmers for each individual application. In this way, severe limitations are built into the systems from the start.

In projects, we start from the unique ability of animals to adapt to different environments. Our approach is to mimic a number of fundamental mechanisms for how animals learn and make decisions. In particular, we mimic the malleability of natural nervous systems, resulting in a dynamic architecture that constantly adapts to new situations. The project is a collaboration between researchers in cognitive science, dynamical systems and biology from Chalmers, Harvard University (USA) and Deakin University (Australia).